Advanced Generalized Predictive Control and Its Application to Tiltrotor Aircraft for Stability Augmentation and Vibration Reduction

Abstract

The goals of this research were to restore generalized predictive control (GPC) capability at NASA and within the community, to better understand GPC and its performance relative to other options, and to improve upon the capability of GPC. Unique to this research is the comparison of GPC with other control options including PID controllers, optimal control theory, and other versions of the similar AutoRegressive moving average model with eXogenous inputs (ARX) models. Similar to GPC, ARX models use an experimentally acquired system identification to characterize the input/output relationship between controls and response measurements. Because this relationship is determined from acquired data, minimal knowledge of the system behavior is required to employ ARX or GPC controllers. As a result of these comparisons, it was observed that GPC is typically the best performing control option and typically has better gain and phase margins when properly employed. Also unique to this dissertation is the use of orthogonal multisine excitation as the command inputs for GPC application rather than the typical distinguishable random noise. Finally, the concept of Advanced GPC (AGPC) is introduced as a part of this dissertation work. AGPC is a self-adapting algorithm that improves traditional GPC when conditions change from those used to derive the system identification. AGPC is also better performing than traditional GPC in some cases even when the conditions do not change from those used to acquire the system identification. Application of AGPC requires the monitoring of performance figures of merit, and the application of control dither when the metrics indicate that the controls are not distinguishable enough or the response of the system is inadequate to properly characterize the input/output relationship. Finally, for experimental application of GPC and AGPC, techniques were introduced to increase model safety and include features such as a magnitude ramp rate when closing the control loop, master gain values to reduce control or dither authority, continual computation of figures of merit, the ability to gradually change from one control algorithm to another, and visualization of control commands prior to closing the control loop and/or switching from one control algorithm to another.